JOURNAL OF MAGNETIC RESONANCE IMAGING 41:1682–1688 (2015)

Original Research

Prediction of Myelopathic Level in Cervical Spondylotic Myelopathy Using Diffusion Tensor Imaging Shu-Qiang Wang, PhD,1,2 Xiang Li, MD,1 Jiao-Long Cui, MD,1 Han-Xiong Li, PhD,3 Keith D.K. Luk, MBBS,1 and Yong Hu, PhD1* Purpose: To investigate the use of a newly designed machine learning-based classifier in the automatic identification of myelopathic levels in cervical spondylotic myelopathy (CSM). Materials and Methods: In all, 58 normal volunteers and 16 subjects with CSM were recruited for diffusion tensor imaging (DTI) acquisition. The eigenvalues were extracted as the selected features from DTI images. Three classifiers, naive Bayesian, support vector machine, and support tensor machine, and fractional anisotropy (FA) were employed to identify myelopathic levels. The results were compared with clinical level diagnosis results and accuracy, sensitivity, and specificity were calculated to evaluate the performance of the developed classifiers. Results: The accuracy by support tensor machine was the highest (93.62%) among the three classifiers. The support tensor machine also showed excellent capacity to identify true positives (sensitivity: 84.62%) and true negatives (specificity: 97.06%). The accuracy by FA value was the lowest (76%) in all the methods. Conclusion: The classifiers-based method using eigenvalues had a better performance in identifying the levels of CSM than the diagnosis using FA values. The support tensor machine was the best among three classifiers. Key Words: cervical spondylotic myelopathy; spinal cord; diffusion tensor imaging; eigenvalue; fractional anisotropy; machine learning J. Magn. Reson. Imaging 2015;41:1682–1688. C 2014 Wiley Periodicals, Inc. V

1 Department of Orthopaedics and Traumatology, Li Ka Shing Faculty of Medicine, the University of Hong Kong, Hong Kong, China. 2 Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, Shenzhen, China. 3 Department of Systems Engineering and Engineering Management, City University of Hong Kong, Hong Kong, China. The first two authors contributed equally to this work. Contract grant sponsor: Research Grants Council of Hong Kong; Contract grant number: HKU 774211M. *Address reprint requests to: Y.H., Dept. of Orthopeadics and Traumatology, University of Hong Kong, 12 Sandy Bay Road, Pokfulam, Hong Kong, China. E-mail: [email protected] Received March 21, 2014; Accepted July 2, 2014. DOI 10.1002/jmri.24709 View this article online at wileyonlinelibrary.com. C 2014 Wiley Periodicals, Inc. V

CERVICAL SPONDYLOTIC MYELOPATHY (CSM) is the most common type of spinal cord dysfunction in patients older than 55 years of age, and the most common cause of acquired spastic paraparesis in the middle and later years of life (1). CSM is the result of narrowing of the cervical spinal canal by degenerative and congenital causes. Surgical treatment is recommended for patients with moderate to severe function deficit and compatible imaging findings (2), and the level of diagnosis is pivotal for surgical planning (2). Neurologic level diagnosis in cervical myelopathy has been employed in performing surgery relatively early (within 1 year of symptom onset), but neurological level diagnosis is considered complicated and difficult in clinical practice (2). Magnetic resonance imaging (MRI) is now widely used for evaluating spinal cord parenchyma. However, conventional MRI, such as T1and T2-weighted imaging, is limited to providing macroscopic information, including gross deformity and hemorrhage (3). Recently, diffusion tensor imaging (DTI) has allowed detection of the microarchitecture of tissue based on a rank-two diffusion tensor model (4). DTI and fiber tractography have advanced the scientific understanding of numerous neurological and psychiatric disorders (5). The most common parameters employed in delineating spinal cord tissue microarchitecture include fractional anisotropy (FA), mean diffusivity, and apparent diffusion coefficient. All of these parameters are derived from eigenvalues to evaluate the scalar properties of water molecule diffusion (6). Eigenvectors and eigenvalues derived from the diffusion tensor matrix reflect the direction and strength of the movement of water molecules (6). Routine T1/ T2 MRI techniques only provide macroscopic level information, while DTI parameters are more sensitive in showing microstructural abnormalities in cervical myelopathy (7). There is a growing interest in the application of DTI to evaluate the spinal cord microarchitecture. For example, Cui et al (8) employed entropy-based principal eigenvector for the evaluation of microstructural changes after cervical myelopathy. Facon et al (9) also reported that DTI can detect myelopathic cord with higher sensitivity and specificity compared with the conventional anatomical MR

1682

Prediction of Myelopathic Level

images. Moreover, Uda et al (10) demonstrated a decrease in FA in most patients with cervical spondylosis. To date, machine learning techniques have been applied to a range of MRI modalities in an effort to automate the diagnosis of mild cognitive impairment and Alzheimer’s disease (11,12). However, few studies have examined the potential for DTI in conjunction with machine learning algorithms to automate the identification of a myelopathic spinal cord. Achieving the automatic classification of healthy levels and myelopathic levels should facilitate the level diagnosis of CSM and clinical determination of surgical strategy.

MATERIALS AND METHODS Subjects This study was approved by the Institutional Ethics Committee. Written informed consent forms were signed by all subjects prior to participation in this study. A total of 74 volunteers, including 58 healthy people and 16 CSM patients age 21 to 84 years, were recruited. The mean age of the healthy group was 40.1 with 31 males and 27 females, and the mean age of the CSM group was 67.7 with 9 males and 7 females. All volunteers were screened to confirm their eligibility. The inclusion criteria of healthy subjects were intact sensory and motor function evaluated by the Japanese Orthopaedic Association score system (13), and negative Hoffman’s sign under physical examination. Exclusion criteria included the presence of neurological signs and symptoms or a past history of neurological injury, diseases, and operations. All the recruited patients were confirmed for a diagnosis of CSM by senior spine surgeons (two surgeons with 30 years of experience and 10 years of experience, respectively). The clinical diagnosis of CSM in this study was based on the neurological examination (two experienced surgeons with 30 years of experience and 10 years of experience, respectively) and imaging findings (image analysis by X.L. with 3 years of experience in DTI image analysis and Y.H. with 10 years of experience in DTI image analysis). The inclusion criteria included 1) numbness or paresthesias in the upper extremities; 2) sensory changes in the lower extremities; 3) motor weakness in the upper or lower extremities; 4) gait difficulties; 5) myelopathic or "upper motor neuron" findings (ie, spasticity, hyperreflexia, clonus, Babinski and Hoffman signs, and bowel and bladder dysfunction); 6) cervical spondylosis and cord compression on conventional MRI. The exclusion criteria were patients with ossification of posterior longitudinal ligament, ossification of ligamentum flavum, congenital stenosis, and other acquired compressive pathology (eg, tumor and calcification) as well as other neurological disorders (eg, multiple sclerosis, amyotrophic lateral sclerosis, and peripheral neuropathy) (14). Clinical Level Diagnosis We employed the index developed by Seichi et al (2) to define the topography of sensory disturbance, levels

1683

of segmental motor innervations, and localization of the reflex center, and made level diagnosis from sensory disturbance, tendon reflexes, and Manual Muscle Test, respectively. Sensory disturbance was defined as at least one of the following three: patientperceived numbness or sensory disturbance detected by light touch or by pinprick. Due to the possible symptom overlap of a higher affected level with the lower ones, the neurological signs may only be able to detect the highest impaired level, and sometimes one or two severely impaired levels underneath. So we combined the result from sensory, motor, and reflex and made the clinical level diagnosis. Imaging Methods Imaging was conducted with a Philips Achieva 3.0T MR system (Best, The Netherlands). During the acquisition process, the subject was placed supine with the SNV head and neck coil enclosing the cervical region, and was instructed not to swallow to minimize motion artifacts. The subject was then scanned with anatomical T1-weighted (T1W), T2-weighted (T2W) imaging, DTI. Sagittal and axial T1W and T2W images were acquired for each subject using a fast spin-echo sequence. The parameters employed in sagittal imaging include: field of view (FOV) ¼ 250  250 mm, slice gap ¼ 0.3 mm, slice thickness ¼ 3 mm, fold-over direction ¼ feet/head, number of excitation (NEX) ¼ 2, resolution ¼ 0.92  1.16  3.0 mm3 (T1W) and 0.78  1.01  3.0 mm3 (T2W), reconstruction resolution ¼ 0.49  0. 49  3.0 mm3, and echo time / repetition time (TE/TR) ¼ 7.2/530 msec (T1W) and 120/ 3314 msec (T2W). A total of 11 sagittal images covering the whole cervical spinal cord were acquired. The parameters used in axial imaging were: FOV ¼ 80  80 mm, slice thickness ¼ 7 mm, slice gap ¼ 2.2 mm, fold-over direction ¼ anterior/posterior (AP), number of excitations (NEX) ¼ 3, resolution ¼ 0.63  0.68  7.0 mm3 (T1W) and 0.63  0.67  7.0 mm3 (T2W), reconstruction resolution ¼ 0.56  0.56  7.0 mm3 (T1W) and 0.63  0.63  7.0 mm3 (T2W), and TE/ TR ¼ 8/1000 msec (T1W) and 120/4000 msec (T2W). Cardiac vector cardiogram (VCG) triggering was used to minimize the impact of the pulsation artifact from cerebrospinal fluid. Image acquisition began as soon as the rise of the wave of QRS complex. A total of 12 transverse images covering the cervical spinal cord from C1 to C7 were acquired, each of which was placed at the center of either a vertebra or an intervertebral disk. Diffusion encoding was performed in 15 noncollinear and noncoplanar diffusion directions with b-value ¼ 600 s/mm2. The parameters employed in imagine acquisition were: FOV ¼ 80  80 mm, image matrix, 128  128, slice thickness ¼ 7 mm, slice gap ¼ 2.2 mm, fold-over direction ¼ AP, NEX ¼ 3, resolution ¼1  1.26  7.0 mm3, reconstruction resolution ¼ 0.63  0.63  7.0 mm3 and TE/TR ¼ 60 msec/5 heartbeats. The image slice planning was the same as in the anatomical axial T1W and T2W images, with 12 slices covering the cervical spinal cord from C1 to C7. Due to the natural curvature of the cervical

1684

Wang et al.

Figure 1. The representative images showing sagittal T2W, B0, three principal eigenvector images (e0, e1, e2) and FA in the healthy cord (A, A0, Ae0, Ae1, Ae2, AFA) and myelopathic cord (B, B0, Be0, Be1, Be2, BFA). The ROI was defined by the B0 image to cover the spinal cord.

spine, it is not possible to set one stack of DTI scans with every scan slice at vertical with the course of the cord. In this study, we used three stacks to fit the curvature of the cervical spinal cord while the DTI scan in each stack lasted about 8 minutes. The average duration of the whole DTI scan was 24 minutes per subject, with an average heart rate of 60 beats per minute. To reduce the impact of the fold-over, spatial saturation with spectral presaturation with inversion recovery was employed. The distortion correction method based on reversed gradient polarity and parallel imaging was used to reduce the EPI distortion impact caused by increased magnetic susceptibility at 3.0T (15,16). DTI Processing In raw DTI images, diffusion-weighting gradients can lead to eddy currents, which results in artifacts. Such artifacts may include shear, false fiber tracking, enhanced background, image intensity loss, and image blurring. These distortions are different for different gradient directions. The goal of DTI processing is to correct the gradient table for slice prescription and correct images for any residual eddy current distortions and motion artifacts using a nonlinear 2D registration and a 3D rigid body registration. In this study the Automated Image Registration (AIR) program (a source code embedded in DTI Studio software, v. 2.4.01 2003; Johns Hopkins Medical Institute, Johns Hopkins University, Baltimore, MD) was employed to reduce the effect of artifact. The realigned and coregistered diffusion-weighted datasets were double-checked for image quality, and then used for estimation of diffusion tensors, including three eigenvalues and the corresponding eigenvectors (for further information about DTI processing, see Soares et al (17)). The region of interest (ROI) was defined by B0 images to cover the spinal cord (Fig. 1). The FA values were calculated and averaged over all selected

voxels in the cord for all subjects using ImageJ (National Institutes of Health, Bethesda, MD). Machine Learning Methods Given the DTI data, the identification of myelopathic levels can be defined as a bi-classification problem. This problem can be solved by introducing some machine learning-based classifiers. The logic behind the method of identifying myelopathic levels is illustrated in Fig. 2. The following classification methods are considered in this task: naive Bayesian (18), SVM (19,20), and STM (21). Considering that both naive Bayesian and SVM are relatively mature methods, we focused on the STM, as follows. In STM, the discriminant function can be given by (21): " # M Y y ¼ sign X wk þ b [1] k¼1

where wk indicates the weight tensor and b the offset. In this study, X is defined as a second-order tensor with the mode-1 fiber indicating the eigenvalues of the diffusion tensor and mode-2 fiber as the region of interest (the dorsal, lateral, and ventral region of both left and right sides). The discriminant function can be described by:   y ¼ sign wT 1 Xw2 þ b [2] where w and b can be calculated by solving the following constrained optimization problem:  N    1  2 X  2 min J wk k¼1 b j ¼   wk  2 þ c ji 2 k¼1 wk j2k¼1 b j Fro i¼1  T  s:t: yi w 1 Xw2 þ b  1  ji ji  0; 1  i  N ; [3] where j is introduced as the slack variable to deal with noise in DTI data.

Prediction of Myelopathic Level

1685

Figure 2. The framework of identifying myelopathic levels using a machine learning-based classifier model. In the training step, the labeled DTI data is employed to classifiers. The machine learning algorithms employed in this work are naive Bayesian, support vector machine, and support tensor machine.

By introducing the positive Lagrange multipliers a and l, the above optimization problem can be rewritten as follows: max

min

l; a wk jM b j k¼1

Lðwk jM k¼1 ; b; j; l; aÞ

[4]

with    1  2  Lðwk jM k¼1 ; b; j; l; aÞ ¼   wk  2 k¼1

2

þc

Fro



N X

N X

ji

i¼1

fold cross-validation (23). In this study, we employed two methods for evaluating the classifiers: 10-fold cross-validation and holdout validation (the traditional validation method). In K-fold cross-validation, the subset size, n, can be optimized by the following steps: i. Divide the data into K roughly equal parts; ii. For each k ¼ 1, 2, . . .. K, fit the model with parameter n to the other K-1 parts, and calculate its error in predicting the kth part. This gives the crossvalidation error:

  li ðyi wT 1 Xw2 þ b

sðn Þ¼

i¼1

 1 þ ji Þ 

N X

K 1X ek ðnÞ K k¼1

[5]

ai ji

i¼1

where ai  0 k1  0 (1  i  N). We compared the results of machine learning methods to clinical level diagnosis and calculated accuracy, sensitivity, and specificity to evaluate the performance of the employed classifiers. Accuracy is calculated by (TPþTN)/(TPþTNþFNþFP), where TP ¼ True Positive, TN ¼ True Negative, FP ¼ False Positive, and FN ¼ False Negative. Sensitivity is defined as TP/(TPþFN) and Specificity is defined as TN/(FPþTN).

Model Evaluation: Cross-Validation The problem with evaluating a proposed model is that it may demonstrate adequate prediction capability on the training data, but might fail to predict future unseen data. Cross-validation is a procedure for estimating the generalization performance in this context (22). In this study, there were two goals for cross-validation: 1. To estimate performance of the learned model from available data using one algorithm; ie, to gauge the generalizability of an algorithm. 2. To compare the performance of naive Bayesian, SVM, and STM, and determine the best algorithm for the available data. In typical cross-validation, the training and validation sets must crossover in successive rounds such that each data point has a chance of being validated. The basic form of cross-validation is k-fold cross-validation. Other forms of cross-validation are special cases of kfold cross-validation or involve repeated rounds of k-

iii. Do this for many values of n and choose the value of n that makes s(n) smallest.

RESULTS Identification of Myelopathic Level Using Naive Bayesian, SVM, and STM To perform holdout validation, the DTI dataset from 20 normal people and 16 CSM patients was divided into two parts. The data from 12 normal people and 8 CSM patients were used for training classifiers, while the data from 8 normal subjects and 8 CSM patients were used for validation (Table 1). The class labels were from neurological diagnosis by senior spine surgeons. In this study the level with CSM was defined as positive and the healthy level as negative. The neurology result in Table 1 indicates the confirmed diagnosis by senior spine surgeons, which was the benchmark in our study. The subjects from case 1 to case 8 are with cervical spinal stenosis and the subjects from case 9 to case 16 are normal people. The dash in Table 1 indicates that there is no level with CSM. The subject of case 13 has no CSM level in the view of the senior spine surgeons. But the classifier based method identified C3–4 as a CSM level. This was a false positive given by the classifier. The statistical results of identification of myelopathic level from the three classifiers are shown in Table 2. The accuracy by STM was the highest (93.62%) of the three classifiers. STM also showed excellent capacity to identify true positives (sensitivity: 84.62%) and true negatives (specificity: 97.06%). Next,

1686

Wang et al.

Table 1 Results of Identification of Myelopathic Level From Neurology, SVM, Bayesian, and STM and FA Value Identification of myelopathic level Case no.

Gender

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16

F M F M F M F M M M F M F F M M

Age 79 83 62 65 43 80 63 60 29 30 49 26 59 52 50 54

Neurology C34, C34, C34, C34 C34 C34, C34, C56 — — — — — — — —

C45 C56 C45

C45 C45

Bayesian

SVM

STM

C45, C56 C34, C45, C56 C34, C56 C34, C56 C45, C56 C45, C56 C45 C34, C56 — — — — C34,C45 — C56 —

C34, C56 C45, C56 C45, C67 C34,C45 C56 C34, C56 C45,C56 C45,C56 — — — — C34 — — —

C34, C34, C34, C34, C34, C34, C56, C56, — — — — C34 — — —

C45, C56 C45, C56 C45, C45, C67 C67

FA C56 C56 C67 C56,

— — C45,C78 C34,C45, C67 C34,C45 C23, C45,C56,C67 C67 — — — — — — — C34, C56 —

Dash indicates there is no level with CSM for the subject.

to compare the performance of naive Bayesian, SVM, and STM, we used the 10-fold cross-validation method. The advantage of this method over repeated random subsampling is that all observations are used for both training and validation, and each observation is used for validation exactly once. The statistical results of identifying myelopathic level with 10-fold cross-validation are shown in Table 3, where STM performed the best among the three classifiers (accuracy: 94.54%; sensitivity: 89.75%; specificity: 98.85%). Comparison With FA Identification of Myelopathic Level FA is one of the most common parameters in DTI. FA is calculated from the eigenvalues of the diffusion tensor, with values between 0 (perfectly isotropic diffusion) and 1 (the hypothetical case of an infinite cylinder). We used FA values from 50 healthy subjects to create a threshold to detect the level with CSM. The means and standard variations of the FA for each level were: C23: 0.7286 6 0.0602, C34: 0.6754 6 0.0657, C45: 0.6877 6 0.0716, C56: 0.6420 6 0.0755, C67: 0.6518 6 0.0525 and C78: 0.6471 6 0.0767. We defined threshold of low FA value (LFV) as LVF ¼ (Mean FA value – 2.5 * SD). The level with FA value below the threshold was defined as a myelopthatic level. The FA values from the first eight cases in Table 1 were used for myelopathic level diagnosis. The identified levels with CSM are listed in column 8 in Table 1, with the statistical results shown in row 5 (accuracy: 76.0%; sensitivity: 30.77%; specificity: 91.89%). The experimental results demonstrate that the classifiers using eigenvalues had a better ability to identify the levels with CSM than the level diagnosis by FA values. DISCUSSION In this study we proposed a data driven-based method to identify the spinal cord levels with CSM.

The eigenvalues of the DTI data were used to train the proposed classifiers, and we compared the eigenvalue-based machine learning method with the FA values. We found that the machine learning-based classifiers were excellent for identifying the levels with CSM in spinal cord. FA is one of the most commonly used indices in DTI analysis (6,24). However, although Uda et al (10) demonstrated a decrease in FA in most patients with CSM, the use of FA only is insufficient to detect the levels with CSM. The current data suggest that the eigenvalues from DTI data can provide more useful information in identifying the levels with CSM in spinal cord than for FA. CSM is a degenerative disease of the cervical spine, which is usually an extensive range of lesions involving multiple segments. Multilevel-affected CSM is complex, with clinical manifestations and it is difficult to precisely localize all the involved levels by neurological examination. However, not all the myelopathic levels appear with high signal intensity on MRI (25). The myelopathic levels identified by MRI findings are usually mismatched with those from neurological examination (2,25). The surgical outcome of CSM still varies a lot and cannot be predicted precisely by either neurologic deficit or any existing imaging method. Therefore, from spinal cord compression to functional deficit, there is a gap, which is the pathological change of spinal cord tissue, which is undetectable by conventional methods but not DTI images. In clinical practice, level diagnosis is critical in determining decompression levels and the surgical Table 2 Statistical Results of Identification of Myelopathic Level Method

Accuracy

Sensitivity

Specificity

Bayesian SVM STM FA

80.85 % 82.48% 93.62% 76.0%

61.54% 53.85% 84.62% 30.77%

88.24% 94.12% 97.06% 91.89%

Prediction of Myelopathic Level

1687

Table 3 Statistical Results of Identification of Myelopathic Level With 10-Fold Cross-Validation Method

Accuracy

Sensitivity

Specificity

Bayesian SVM STM

83.51 % 83.27% 94.54%

66.47% 59.39% 89.75%

90.31% 96.48% 98.85%

approach (anterior or posterior). In the case of multilevel CSM, conventional MRI may detect some mild compression levels beside the severest compressed level. If the neurological deficit does not clearly point to the mildly compressed levels, it will be controversial for surgical planning. Surgical decisions made purely on conventional MRI will put the patients at the risk of "overkill" or inadequate decompression. DTI could reveal the microstructure impairment and reflect the pathological changes of the spinal cord based on its unique principle. With the method introduced in the present study, the myelopathic levels could be identified in CSM patients based on detection of the demyelination of white matter. It could reveal the pathological condition of an impaired cervical spinal cord in an efficient way. In the present study, STM was especially suitable for identification of spinal cord levels with CSM. As a relatively new learning approach, STM has some potential advantages in dealing with DTI data. While traditional linear classification algorithms like SVM find a classifier in Rm, STM finds a classifier in tensor space Rm1  Rm2, which provides a structured classification. Therefore, STM can use the DTI data structure, while SVM often results in data structure loss when translating a tensor into a vector. The number of independent unknown parameters in STM is also less than that of SVM. For example, a vector X 僆 Rn can be transformed to a second-order tensor X 僆 R n1  R n2 , where n  n1  n2. In SVM, a linear classifier can be represented as kTxþb in which there are nþ1 ( n1  n2þ1) independent parameters (b, ki, i ¼ 1, 2. . .n). In STM, a linear classifier can be represented as x1TXx2þb where x1僆Rn1 and v1僆Rn2. Thus, there are only n1þn2þ1 parameters. This property makes STM especially suitable for small sample cases and is robust against overfitting. Prior to the use of advanced imaging techniques, neurological examination was the main approach to estimate the level of myelopathy, and remains an essential method for evaluating the severity and location of the lesion. The cervical cord segments approximately correspond to one or two higher intervertebral levels in CSM (2), owing to the different anatomical relationship between cord segments and spinal roots with regard to intervertebral levels. Although it is difficult to distinguish all the myelopathic versus normal levels in cases of multilevel involvement, neurologic examination can provide the most direct evidence that certain levels of the cervical spinal cord exhibit myelopathy. Thus, neurological examination remains a benchmark for identifying cervical spinal cord myelopathy for comparisons to DTI methods.

In this study, we demonstrated that the machine learning-based classifiers using eigenvalues of DTI can provide a direct measure of the level of myelopathy, with the STM-based classifier providing the optimal detection method. There are several advantages of the STM-based classifier in identifying the levels with CSM in the spinal cord. First, compared with the vector space model, STM can exploit the DTI data structure, as well as correlations in the original data. Thus, the STM-based classifier allows the detection of CSM levels with higher accuracy and sensitivity. Second, the number of independent parameters in STM is less than that of the vector-based classifier, which makes the STM-based classifier more robust against overfitting compared with the vector space model, such as the SVM-based classifier. This also allows the STM-based classifier to deal with small sizes, which are very common in medical science. Finally, the STM-based classifier is more cost-effective than neurological examination by senior spine surgeons, and this classifier can work efficiently as long as it is well trained. Therefore, the proposed STM-based classifier would provide additional diagnosis of myelopathic levels for surgeons to make the most appropriate surgical plan. Several issues in the STM-based classifier should be considered. The first is how to sort the features in the tensor. In SVM, we implicitly assume that the features are independent (19). A classifier in vector space can be written as kTXþb. Obviously, the change of the order of the features has no impact on training the classifier. In the tensor space model, a linear classifier is represented as VTXUþb. Therefore, the independence assumption for features no longer holds in training the tensor-based classifier, and different feature sorting will lead to different training results in the tensor space model. In this study, we sorted the features into descending order of eigenvalues. The second issue is the loss of features caused by the ROI. Note that the eigenvalue employed in this study is the average of those from the voxels drawn within the ROI. Thus, the definition of the ROI is important. Unsuitable drawing of the ROI may lead to feature loss in training the classifier. In future studies, we will consider training the classifier using DTI data from all voxels of the whole spinal cord. Finally, identifying the levels with CSM is only the first step. Estimating the pathological severity at each level is necessary in clinical practice. There are some limitations to the present study. Since CSM has a vast array of signs and symptoms and there are no pathognomonic findings, selection bias may happen in the inclusion of patients by experienced surgeons. Also, different disease severity and compression patterns in multilevel cases may also bias the result. Another limitation of the present study is the lack of neurological examination results from CSM patients after treatment, which may provide sufficient information on justifying the classifiers. Although neurology could not reveal the precise information of myelopathy along cervical cord segments, it is the only available and acceptable benchmark after a careful clinical diagnosis by experienced spine

1688

surgeons. A more appropriate reference standard should be based on the surgical outcome following level diagnosis, which needs a large-scale clinical trial in a future study. In conclusion, the proposed machine learningbased method might provide a valuable method for predicting the changes of clinical symptoms and the estimated pathological severity at each level over time. With the proposed classifiers, we could detect the myelopathic levels in CSM and provide useful reference to spine surgeons in decision making of a surgical plan in complicated cases. REFERENCES 1. Montgomery D, Brower R. Cervical spondylotic myelopathy. Clinical syndrome and natural history. Orthop Clin N Am 1992;23: 487–493. 2. Seichi A, Takeshita K, Kawaguchi H, et al. Neurologic level diagnosis of cervical stenotic myelopathy. Spine 2006;31:1338–1343. 3. Baron EM, Young WF. Cervical spondylotic myelopathy: a brief review of its pathophysiology, clinical course, and diagnosis. Neurosurgery 2007;60(1 Suppl 1):S35–41. 4. Thurnher M, Law M. Diffusion-weighted imaging, diffusion-tensor imaging, and fiber tractography of the spinal cord. Magn Reson Imaging Clin N Am 2009;17:225–244. 5. Mukherjee P, Berman J, Chung S, Hess C, Henry R. Diffusion tensor MR imaging and fiber tractography: theoretic underpinnings. AJNR Am J Neuroradiol 2008;29:632–641. 6. Hagmann P, Jonasson L, Maeder P, Thiran JP, Wedeen VJ, Meuli R. Understanding diffusion MR imaging techniques: from scalar diffusion-weighted imaging to diffusion tensor imaging and beyond. Radiographics 2006;26(Suppl 1):S205–223. 7. Demir A, Ries M, Moonen C, et al. Diffusion-weighted MR imaging with apparent diffusion coefficient and apparent diffusion tensor maps in cervical spondylotic myelopathy. Radiology 2003;229:37–43. 8. Cui J-L, Wen C-Y, Hu Y, Li T-H, Luk K. Entropy-based analysis for diffusion anisotropy mapping of healthy and myelopathic spinal cord. Neuroimage 2011;54:2125–2131. 9. Facon D, Ozanne A, Fillard P, Lepeintre J-F, Tournoux-Facon C, Ducreux D. MR diffusion tensor imaging and fiber tracking in spinal cord compression. AJNR Am J Neuroradiol 2005;26:1587– 1594.

Wang et al. 10. Uda T, Takami T, Tsuyuguchi N, et al. Assessment of cervical spondylotic myelopathy using diffusion tensor magnetic resonance imaging parameter at 3.0 Tesla. Spine 2013;38:407–414. 11. Desikan R, Cabral H, Settecase F, et al. Automated MRI measures predict progression to Alzheimer’s disease. Neurobiol Aging 2010;31:1364–1374. 12. Magnin B, Mesrob L, Kinkingn ehun S, et al. Support vector machine-based classification of Alzheimer’s disease from wholebrain anatomical MRI. Neuroradiology 2009;51:73–83. 13. Yonenobu K, Abumi K, Nagata K, Taketomi E, Ueyama K. Interobserver and intraobserver reliability of the Japanese Orthopaedic Association scoring system for evaluation of cervical compression myelopathy. Spine 2001;26:1890. 14. McCormick WE, Steinmetz MP, Benzel EC. Cervical spondylotic myelopathy: make the difficult diagnosis, then refer for surgery. Cleveland Clin J Med 2003;70:899–904. 15. Chuang T-C, Huang T-Y, Lin F-H, et al. PROPELLER-EPI with parallel imaging using a circularly symmetric phased-array RF coil at 3.0 T: application to high-resolution diffusion tensor imaging. Magn Reson Med 2006;56:1352–1358. 16. Morgan P, Bowtell R, McIntyre D, Worthington B. Correction of spatial distortion in EPI due to inhomogeneous static magnetic fields using the reversed gradient method. J Magn Reson Imaging 2004;19:499–507. 17. Soares J, Marques P, Alves V, Sousa N. A hitchhiker’s guide to diffusion tensor imaging. Front Neurosci 2013;7:31. 18. Michael DP. On the optimality of the simple Bayesian classifier under zero-one loss. Mach Learn 1997;29:103–137. 19. Cortes C, Vapnik V. Support-vector networks. Machine Learn 1995;20:273–297. 20. Steinwart I, Christmann A. Support vector machines. New York: Springer; 2008. 21. Kotsia I, Guo W, Patras I. Higher rank support tensor machines for visual recognition. Pattern Recog 2012;45:4192–4203. 22. Picard R, Cook D. Cross-Validation of Regression Models. J Am Stat Assoc 1984;79:575–583. 23. Stone M. Asymptotics for and against cross-validation 1977;64: 29–35. 24. Budzik JF, Balbi V, Le Thuc V, Duhamel A, Assaker R, Cotten A. Diffusion tensor imaging and fibre tracking in cervical spondylotic myelopathy. Eur Radiol 2011;21:426–433. 25. Zhang L, Zeitoun D, Rangel A, Lazennec JY, Catonn e Y, PascalMoussellard H. Preoperative evaluation of the cervical spondylotic myelopathy with flexion-extension magnetic resonance imaging: about a prospective study of fifty patients. Spine 2011; 36:9.

Prediction of myelopathic level in cervical spondylotic myelopathy using diffusion tensor imaging.

To investigate the use of a newly designed machine learning-based classifier in the automatic identification of myelopathic levels in cervical spondyl...
266KB Sizes 0 Downloads 3 Views